logo
ResearchBunny Logo
Instantaneous tracking of earthquake growth with elastogravity signals

Earth Sciences

Instantaneous tracking of earthquake growth with elastogravity signals

A. Licciardi, Q. Bletery, et al.

This groundbreaking research by Andrea Licciardi, Quentin Bletery, Bertrand Rouet-Leduc, Jean-Paul Ampuero, and Kévin Juhel introduces prompt elastogravity signals (PEGS) and a deep learning model, PEGSNet, to revolutionize real-time earthquake magnitude estimation for large earthquakes. Discover how this innovation enhances tsunami early warning capabilities by tracking earthquake growth before seismic waves arrive.

00:00
00:00
Playback language: English
Introduction
Accurate and rapid estimation of large earthquake magnitudes (above 8) is crucial for mitigating risks associated with strong shaking and tsunamis. Current early warning systems relying on seismic waves often fail to provide timely and accurate estimates for these large events. Geodesy-based approaches offer better estimations but suffer from latency and uncertainty due to the speed of seismic waves. Recent studies have shown the potential of speed-of-light prompt elastogravity signals (PEGS) to overcome these limitations. However, operational testing of PEGS for early warning has been lacking. This research addresses this gap by demonstrating the use of PEGS for real-time tracking of earthquake growth. The study's importance lies in its potential to revolutionize tsunami early warning by providing nearly instantaneous information on the evolving magnitude of a large earthquake, before seismic waves arrive, giving precious time for preparation and evacuation.
Literature Review
Existing earthquake early warning (EEW) systems based on seismic waves struggle with accurate magnitude estimation for large earthquakes (M>8). These systems often saturate, underestimating the final magnitude. Geodesy-based approaches, using Global Navigation Satellite System (GNSS) data, have emerged as an alternative, offering better magnitude estimations. However, these approaches are not without drawbacks, including uncertainties related to data selection and preprocessing, and questions remain regarding the true predictive power, particularly regarding the ability to estimate final magnitude before rupture completion. Deep learning has shown promise in seismology, improving the detection of small seismic signals and characterizing earthquake source parameters. This paper builds upon this progress, applying deep learning to PEGS data for enhanced EEW.
Methodology
The study focuses on the Japanese subduction zone due to its history of large subduction events and dense network of high-quality seismic stations. Due to limited real PEGS observations, the researchers trained their model, PEGSNet, using a large database of synthetic PEGS waveforms generated using a normal-mode summation code. Real noise recorded at various seismic stations was added to the synthetic data to improve model robustness. The training database comprised 500,000 synthetic earthquake sources with varying parameters (location, magnitude, strike, dip, rake angles) based on the Slab2.0 model. Simplistic point-source rupture descriptions were used initially, although more complex STF models were also considered. PEGSNet, a convolutional neural network (CNN), was designed to estimate earthquake location and track the time-dependent magnitude M<sub>w</sub>(t) from PEGS data. The data was preprocessed by bandpass filtering (2 mHz - 30 mHz) and clipping to enhance PEGS amplitudes and suppress noise. The input data for PEGSNet was arranged as a three-channel image (one for each seismic component), with stations sorted by longitude. A novel training strategy involved randomly selecting the starting time of the analyzed data window, allowing the model to learn patterns in the data as the source time function (STF) evolves. P-wave arrival times were assumed known, enabling the model to exclusively use pre-P-wave PEGS data. The model was trained to predict M<sub>w</sub>(t), latitude, and longitude, using the Huber loss function. The model was tested using a separate test set of synthetic data and then applied to the real data from the 2011 Tohoku-Oki earthquake.
Key Findings
PEGSNet demonstrated high accuracy in tracking moment release for earthquakes with final M<sub>w</sub> above 8.6, achieving accuracy above 90% and errors below 0.25 magnitude units, starting approximately 40 seconds after the origin time. For earthquakes with final M<sub>w</sub> between 8.2 and 8.6, accurate estimation was possible after around 150 seconds. A conservative lower limit on PEGSNet's sensitivity to M<sub>w</sub> was set at 8.3, although under favorable noise conditions, this could be reduced to 7.9 or 8.0. Analysis of predictions for events with a final M<sub>w</sub> of 9.0 ± 0.05 revealed underestimation in the first 30-40 seconds due to the lower sensitivity limit and potential masking of PEGS by P-waves at near-source stations. Beyond this initial period, PEGSNet provided instantaneous tracking of moment release with minimal time shift between estimated and true M<sub>w</sub>(t). Application to the 2011 Tohoku-Oki earthquake showed PEGSNet's superior performance compared to existing EEW systems (JMA, FinDer, BEFORES) in terms of both latency and accuracy, particularly after approximately 55 seconds (when M<sub>w</sub> reached 8.3). PEGSNet consistently provided estimates closer to the true STF than other algorithms, starting at 55 s and reaching a correct prediction of the final M<sub>w</sub> after about 120 s. Notably, the algorithm did not suffer from magnitude saturation. A test using noise-only data showed that PEGSNet predictions for Mw < 8 converged to the noise baseline. Additional tests on all subduction earthquakes (M<sub>w</sub> ≥ 7) since 2003 (excluding aftershocks) confirmed that events with M<sub>w</sub> < 8 were indistinguishable from noise.
Discussion
The findings demonstrate the potential of PEGS as a new observable for real-time earthquake magnitude estimation. PEGSNet's ability to instantaneously track moment release for large earthquakes addresses the limitations of existing seismic and GNSS-based EEW systems. The speed-of-light propagation of information in PEGS allows for significantly faster and more accurate magnitude estimation, particularly for events above M<sub>w</sub> 8.3. This has direct implications for tsunami early warning, where timely magnitude estimates are crucial for accurate tsunami forecasts and risk mitigation. The model's robustness is supported by its consistent performance on both synthetic and real-world data, including the challenging 2011 Tohoku-Oki earthquake. The ability to use routinely discarded pre-P-wave data makes the approach cost-effective, requiring no additional instrumentation.
Conclusion
This research successfully demonstrated the potential of PEGS for instantaneous tracking of earthquake moment release. The PEGSNet model offers a significant advancement in EEW, particularly for large earthquakes, addressing the limitations of traditional systems. The results highlight the importance of PEGS as a new and independent source of information for real-time magnitude estimation, especially beneficial for tsunami early warning. Future research could explore improvements to the STF model for more accurate estimations, adapt the model for different tectonic settings and seismic networks, and integrate PEGSNet with other existing EEW systems for a comprehensive early warning approach.
Limitations
The study's reliance on a synthetic training dataset, while necessary given the scarcity of real PEGS data, introduces potential limitations. The accuracy of the synthetic data generation process directly impacts the model's performance. Additionally, the use of a simplified point-source rupture model may not fully capture the complexity of real-world rupture processes, potentially affecting the accuracy of the magnitude estimations. Finally, the assumption of known P-wave arrival times needs to be considered in an operational context.
Listen, Learn & Level Up
Over 10,000 hours of research content in 25+ fields, available in 12+ languages.
No more digging through PDFs, just hit play and absorb the world's latest research in your language, on your time.
listen to research audio papers with researchbunny